The split
| AI did | People did |
|---|---|
| Wrote most of the routine code: forms, tables, lists, permissions | Sat with the team and watched how orders really move |
| Generated realistic test data from real document formats | Decided which stages exist and who can change them |
| Drafted screens, help text and the handbook | Chose what each role can see, from director to supervisor |
| Wrote and ran tests on every change | Reviewed every release before anyone used it |
| Suggested structures for offers and numbering | Said no to features that would make it harder to use |
AI is fast and tireless, but it doesn’t know your business. People know the business, but can’t type fast enough to build a 46-screen system at a price a growing company can justify. Together, you get software that fits, in weeks instead of months.
Eight rules we built by
Scope items, operations, stages, payment terms, clauses, option lists, approval rules and numbering series are all rows the business edits itself. Anything that is a list of choices, the customer can change. Only real logic lives in code.
Progress, percentages and what is claimable are worked out from the facts. A number nobody types is a number nobody gets wrong.
If a customer has no client code, an offer can’t be numbered, and the system says so. If a value falls outside every approval band, it is refused, not waved through. Gaps surface as questions, not habits.
Offers are revised, not edited. Customer orders are amended, not overwritten. Old versions are kept, because someone will always ask about “the price you quoted in August”.
Approval by exception, with a written reason, exists because real life sometimes needs material ordered at night. A recorded override beats a workaround.
Hiding a menu item is not security. A supervisor who types the receivables address directly is refused.
Every night the whole database is exported to a readable Excel workbook. If the software ever disappeared, the business would still have everything.
StageVO started on free and low-cost tiers of mainstream services. Running costs grow only when the number of customers does.
How the work actually flowed
- Collect the paper
We started with the documents people already used: offers, purchase orders, job cards, challans, the Excel tracker. They are the real specification.
- Write the rules down
Numbering formats, scope items, payment stages and approval rules were agreed in writing before any screen was built.
- Build the most painful part first
Money: claims, receivables and retention. If that worked, the rest of the system had a reason to exist.
- Put it in front of people early
Supervisors and the office used early versions on live orders. Their corrections changed more than any design review would have.
- Write the handbook as you go
Setup, daily use and troubleshooting were documented alongside the code, so the system doesn’t depend on the people who built it.
What we learned
- Talk to the shop floor early. They know when a stage is really done, often before the office does.
- Clean old data in the first week, not the fifth. Open jobs with retention outstanding are the most valuable rows you will enter.
- Every rule you make a setting saves a support call later.
- AI makes building cheap. It doesn’t make deciding cheap. Spend the time you save on understanding the work.